SpaceX AI outlines two Colossus sites
Two planned Mississippi sites would add training and inference infrastructure to Colossus 2, with approximately 420,000 GPUs envisaged across them.
The next stage of an AI infrastructure plan is often expressed in GPU counts. Turning those counts into useful computing capacity requires facilities that are powered, connected and commissioned.
What happened
SpaceX AI outlined two additional Colossus 2 sites in Southaven, Mississippi, called Gun Range and Spam. The announced plan envisages approximately 420,000 Nvidia GB300 GPUs across the two locations.
Gun Range is intended for AI training, while Spam would convert a former warehouse into an inference facility. The stated operating targets were the end of November and December respectively. These are planned milestones and equipment totals, rather than independently confirmed installed capacity.
Why it matters
Separating training and inference can reflect different workload requirements. Training builds or improves models, while inference runs them to answer requests and complete tasks.
Additional facilities could support both activities, but the usable capacity depends on more than equipment purchases. Power delivery, cooling, networking and reliable operation all have to work together.
The bigger picture
Large announced clusters illustrate the continuing ambition behind AI computing investment. They should also be assessed in stages, distinguishing a site plan from installed equipment and a commissioned service.
For these two locations, the relevant next evidence is deployment and operating progress. The GPU totals and schedules remain targets as of the announcement, with the eventual contribution to the broader Colossus network dependent on execution.